Polarimetric SAR urban classification using the Touzi target scattering decomposition

Abstract

The Touzi decomposition was recently introduced for a roll and unique incoherent decomposition of target scattering. Unlike the Cloude–Pottier incoherent target scattering decomposition, no scattering parameter averaging is performed and target scattering is fully characterized by an in-depth analysis of five parameters for each of the three (dominant, medium, and low scattering) single scattererings. For practical reasons, there is an immediate need for a procedure that permits the ranking and selection of the smallest Touzi decomposition parameter subset that permits an accurate image classification. In this study, a filter method based on mutual information criteria was introduced to select and rank the most suitable decomposition parameter subset for the classification of an urban site in Ottawa, Ontario. Two approaches are considered for the selection of the suitable parameter subset: the maximum mutual information and the maximum relevancy and minimum redundancy. The two approaches were compared in terms of feature selection and class discrimination, and the optimum subsets selected were used to perform an urban site classification based on a support vector machine. Polarimetric Convair 580 synthetic aperature radar data collected over an urban site in Ottawa were used to assess and validate the selection and classification results.